Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

jev-nano

Tiny /v1/systemone-compatible decision server: a 7M-parameter CPU model in an 8 MB native binary, with conformal abstention.

Modelprototype
View on GitHub ↗

Overview

jev-nano is described as the nano tier of local Jev: a /v1/systemone-compatible server answering noul, choice and score questions with probabilities, running a 7M-parameter model on any Linux x86_64 CPU from an 8 MB tarball. It is written in MFL, compiled through the machin C backend into a single static binary with no Python in the inference path. Any Jev or TypeSafe SDK client can point at it via base_url.

It uses a frozen trunk plus small trained heads with candidate log-probability scoring, and a calibrated conformal gate for abstention. The README reports 19 of 19 wire-contract checks passing and extra endpoints for routing with abstention, head lanes, chat completions and feedback. It is candid about quality: on JevBench's public set it scored 28% on the easy and original tiers (hard-tier states exceed its 384-token context), agreeing with hosted Jev on clear noul calls but near chance on fine-grained multi-way choices.

What the decision model decides

TypeQuestionAnswers
Yes / noDoes this convey urgency?yes · no
ChoicePick one option from the criteria map—
ScorePlace the state on the score legend—

How it works

  1. Install with the script and start the binary
  2. Point a Jev SDK client or curl at http://127.0.0.1:8097
  3. Send state and typed questions to /v1/systemone
  4. Get probabilities, or an abstention from the conformal gate

Strengths

  • Runs on any CPU with an 8 MB download
  • Drop-in /v1/systemone wire compatibility
  • No Python or GPU needed
  • Honest public benchmark results

Caveats

  • Low accuracy on multi-way choices (28% on JevBench easy and original tiers)
  • 384-token context limit
  • Written in a niche language (MFL)

Getting started

curl -fsSL https://raw.githubusercontent.com/javimosch/jev-nano/main/install.sh | bash, then ./demo.sh

At a glance

Type
Model
Status
prototype
Language
MFL
Runtime
Native binary HTTP server (Linux x86_64)
Has code
yes
Has tests
yes

Use cases and integrations

How curated lists describe it

Frequently asked questions

Can I use the TypeSafe SDK with it?

Yes. Point the client's base_url at http://127.0.0.1:8097.

How accurate is it?

The README reports 28% on JevBench's easy and original tiers; fine on clear yes/no calls, near chance on multi-way choices.

Does it need a GPU or key?

No. It runs on CPU locally and needs no TypeSafe key.

Related projects

  • brier

    Mangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.

    ProjectproductionPython
  • blink★ 22

    Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.

    ServicebetaC
  • Foq★ 11

    Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.

    ProjectbetaPython
  • browser-laya

    Runs the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.

    ServicebetaTypeScript
  • ifReflex.cpp

    llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.

    ProjectbetaC++
  • jevalt

    Three open 4B decision models for English, Turkish and German that serve the Jev /v1/systemone API on a CPU in about 3 GB of RAM.

    ServicebetaPython